hotfix: plan-time retry on hallucinated skill names

[AVADSA25]

Reproducer (real, today)

User dropped:

"Read all markdown files in ~/codec-repo/docs/ and create an index.md that lists each file with its first heading and a one-line description"

Result:

❌ Plan failed: plan invalid: plan references unknown skills: ['file_read']

The user never even reached approve/reject. Plan validation rejected the draft because Qwen invented `file_read` (the actual skill is `file_ops`).

Why PR #35 didn't catch it

PR #35 added a single-shot correction-nudge retry inside `codec_agent_runner._execute_checkpoint` — that handles hallucinations at execution time (skill / write_path / read_path / domain). But validation lives earlier, in `codec_agent_plan.draft_plan` → `validate_plan_skills`. The plan is rejected before it's even saved, so the runner never gets a chance to retry.

Fix (mirrors PR #35 one layer up)

1. After `validate_plan_skills` returns `ok=False`, build a corrective prompt with:
   • The missing skill names
   • The FULL allowed registry list (closed-world choice)
   • The three most common confusions: `file_read`→`file_ops`, `fetch_url`→`web_fetch`, `read_file`→`file_ops`
2. Re-call Qwen ONCE with the appended correction.
3. Re-validate. Success → use the corrected plan. Failure → raise with both attempts in the message so the user sees consistent confusion (vs a one-off transient miss).
4. If the retry call itself flakes, surface the original validation error.

Also strengthen `_PLAN_SYSTEM_PROMPT` with the same three confusion hints so the FIRST draft is more likely to succeed (cuts the retry rate).

Tests (3 new)

• `test_draft_plan_retries_on_hallucinated_skill_then_succeeds` — reproduces the exact user case (file_read → file_ops)
• `test_draft_plan_retry_also_fails_raises_with_both_attempts` — both attempts miss, message contains both
• `test_draft_plan_retry_qwen_unavailable_surfaces_original_error` — retry call raises ConnectionError, original validation error surfaces

All 3 existing `draft_plan` tests still pass — backward-compat preserved. `test_draft_plan_rejects_unknown_skill` now exercises BOTH attempts (fake returns same bad plan twice) and still raises with the missing skill in the message.

Tally: 35/35 file tests pass + 7 pre-existing pynput env failures (unchanged on baseline, unrelated to this PR).

Test plan after merge

• Drop the markdown-index project again — expected: plan succeeds with `file_ops`, agent runs, index.md written
• Drop the forex briefing again — expected: same fix path also helps if Qwen ever drafts it with hallucinated names

🤖 Generated with Claude Code